This study compares machine-learning algorithms and molecular representations for predicting carbon-dioxide solubility in deep eutectic solvents. It uses 2,648 experimental measurements from 93 hydrogen-bond acceptor–donor systems and reports the best test performance from a structural-code random forest.
Key findings
- The structural-code random forest achieved the strongest reported test performance, with R² = 0.971 under the study’s adopted random split. The result supports structure-based screening within chemical space similar to the training data.
Why this matters globally
With prospective validation on new chemistries and operating conditions, this approach could reduce preliminary experiments in carbon-capture solvent screening. The study does not demonstrate process- or plant-scale capture performance.
Thai researcher contribution
Teerawat Sema of Chulalongkorn University is the corresponding author, showing a Thai institutional role in data-driven modelling for carbon-capture materials and processes.
Limitations to consider
The dataset is dominated by choline-chloride systems, with limited betaine, ammonium-salt and hydrophobic DES coverage. Literature-derived measurements may be heterogeneous. The split does not establish extrapolation to unseen chemistry, and no prospective or process-scale validation is reported.
Verify the original sources
Clean EnergyRead the original article↗DOI: 10.1093/ce/zkag039